Alex Towfigh

Founder & CEO of AquaMesh. Bring physical ai autonomy to factories.

[Alex Towfigh @alextowfigh]

@chris_j_paxton UR's CB3 arms already estimated force from joint current at about ±10 N, and UR still added a wrist force sensor. If blind policies hit 73% zero shot on peg insertion, does proprio replace tactile or just sort which jobs never needed it?

[Alex Towfigh @alextowfigh]

@liu_mingyu Cosmos learned physics from 20M hours of video, but most plants already sit on years of logged process data that never trains anything. Isn't the next industrial world model built from historians, not cameras?

[Alex Towfigh @alextowfigh]

@RonVokoun Trevi needs 60 to 80°C heat, but air cooled halls reject 25 to 35°C and even direct liquid cooling mostly lands at 50 to 60°C. So is the 200k gal/MW/day claim really a bet on running chips hotter?

[Alex Towfigh @alextowfigh]

@Scobleizer @1x_tech @jack Neo's harder chores still route to a remote 1X operator, with one manager watching eight of them. A humanlike voice raises expectations, so does it hide that gap or expose it faster?

[Alex Towfigh @alextowfigh]

@adcock_brett Helix launched on about 500 hours of teleop data and Figure now spends $8.1M on data in a week. Once anyone with capital can buy human video, isn't the real moat hours logged by robots doing paid work?

[Alex Towfigh @alextowfigh]

>frontier models write code, essays and legal memos
>on CAD, a model from Maluuba's founder just beat all of them
>51-125x faster, 150-210x cheaper

words were the easy data. the moat in physical AI is geometry that actually got built

[Sam Pasupalak @spisallyouneed]

Introducing SKY-001, the first Engineering World Model.
On CAD and BIM tasks, it beats GPT-6 Astra and every frontier model we tested:

-> 51-125x faster
- > 150-210x cheaper
-> Higher IOU than all frontier models

SKY-001 lets you dream a physical object before you build it.

18 months of work. Built by the team out of Maluuba's deep learning lab (acq. by @MSRResearch).

The first ‘AI General Engineer’ will be the backbone for Autonomous Manufacturing.

[Alex Towfigh @alextowfigh]

A subwoofer company lost DARPA's robot-car race in 2005. By 2007, 5 of the 6 cars that finished the Urban Challenge, winner included, saw through its spinning 64-laser sensor. Everyone built drivers. It built the eyes. You find the bottleneck by entering the race and losing.

Pics: the sensor spinning on a robot car, and its insides from founder David Hall's patent drawing.

[Alex Towfigh @alextowfigh]

>need low-power cameras for deep-space probes
>CCDs gulp ~100× more power than the craft can spare
>colleagues: "you're an idiot to work on this"
>builds CMOS camera-on-a-chip at JPL anyway
>1995: spins Photobit from a home PC on maternity leave
>Micron buys it. CMOS eats the camera market.

Your phone camera exists because a spacecraft couldn't afford a CCD.

[Alex Towfigh @alextowfigh]

> fingertip tactile sensor
> pressure, texture, vibration, slip, heat
> robot sorts objects with cameras taped shut
> still generalizes to new sizes and firmness

Cameras were never the hard part. Contact was.

[Rhoda AI @RhodaAI]

What can a robot do without vision?

We gave our robot a sense of touch and showed it can recognize and retrieve objects with its cameras blindfolded.

Introducing Rhoda’s new tactile sensor—it feels pressure, texture, vibration, slip, and temperature, just like your fingertips.

[Alex Towfigh @alextowfigh]

Every factory, farm, and city is flying half-blind on its water.

A chemical spike can kill a batch, a fish farm, or a river. And they usually find out days later, from a lab.

So we built a sensor that shines light through water and reads 16+ chemicals at once, live, from one probe. It replaces a $100k rack of separate instruments.

Our first one missed the goal by 100x. Like trying to hear a whisper in a stadium.

Rebuilt it. Still 10x short.

Rebuilt again. Beat the goal by 10x.

Most people quit at the first 100x. We treated every miss as data.

[Alex Towfigh @alextowfigh]

>spend 10 years chasing a wind-powered land speed record
>four failed land yachts
>every credit card but one maxed out
>finally hit 126.2 mph in the Mojave in 2009
>"no real long-term strategy other than record-breaking"

12 years later, his engineless robot boat sailed through the eyewall of a Cat 4 hurricane and logged a 126.4 mph gust.

Same wing tech. Almost the same number.

Nobody funds a decade-long hobby. That's exactly why nobody could copy it.

[Alex Towfigh @alextowfigh]

>50B MoE for the physical world
>pretrained on 1M hours of first-person human video
>plain English → 20-DoF robot hands
>zero-shot wontons, brush writing, 30-min continuous runs

The bottleneck was never the actuators. It was teaching a machine the way a human grows up.

[nature will @naturewill]

Introducing Natura-1, our first general foundation model for the physical world.

It follows natural-language instructions zero-shot and turns them into action with dexterous hands.

Natura-1 is a 50-billion-parameter mixture-of-experts (MoE) model pretrained on one million hours of egocentric human experience.

On top of this foundation, we conduct efficient post-training on a limited amount of robot data to adapt pretrained world knowledge to dexterous manipulation, yielding a model that generalizes to a wide range of unseen scenarios.

Natura-1 follows complex natural-language instructions zero-shot, reasoning through spatial constraints, multi-object relationships, and negative conditions.

It can also autonomously carry out long and continuous workflows to perform long-horizon executions.

Natura-1 provides early evidence that generalization—and even emergent behavior—can coexist with the complexity of dexterous hands.

Further progress depends on the full stack: infrastructure for faster training and real-time inference, pretraining data at greater scale and coverage, and models that transfer to robots with less adaptation.

We expect it to be a long-term research and engineering effort.

[Alex Towfigh @alextowfigh]

@treg_ai Can someone explain how this works from a tech perspective? I would assume a lot of the cost of the current databases is to ensure accuracy.

[Alex Towfigh @alextowfigh]

IF the data is accurate this is could be insane.

One step closer to closing the AI distribution gap.

Tech -> Market is the hardest part

[Treg @treg_ai]

Introducing OpenSource people-search for Claude

- Juts $0.0089/lead, no more $600 subscriptions
- #1 on People Search Bench
- 0% markup, Fully open source

Try it at https://t.co/D2eDbgiLu0

Repo link below 👇

[Alex Towfigh @alextowfigh]

>12m autonomous airship, helium lift, hydrogen fuel cell
>hovers 25m over power lines, pipelines, rail
>10 hrs aloft carrying 10kg of LiDAR, cameras, methane sniffers
>1mm per pixel imaging
>founders avg age 27, certifying an aircraft AND selling to utilities

HyLight just raised €13M. France's grid operator Enedis still plans 120,000 km of helicopter line inspection this year, vs ~15,570 km by airship.

The edge isn't the photo. It's the 50th flight over the same pylon, when a fault shows up as a change instead of a guess.

[Y Combinator @ycombinator]

Congrats to @hylight_aero (S23) on their $14.5M raise!

They’re bringing back the airship, this time autonomous, hydrogen-powered, and built to inspect critical infrastructure.

Power lines, pipelines, and railways are still mostly inspected by helicopter, which is expensive, polluting, and risky for crews. HyLight’s 12-meter airships hover low and slow with LiDAR, cameras, and methane detectors, capturing imagery down to 1 millimeter per pixel. Four are flying missions in France today, with 15 planned to operate at once by 2027.

https://t.co/g4CfPmyQot

[Alex Towfigh @alextowfigh]

>sells 35¢ aquarium pH sensors after borrowing $10k from mom
>wins a DOE grant to measure seawater chemistry
>needs a spectrometer that fits in an underwater housing
>none exist, so they build one
>~1,000× smaller and 10× cheaper than lab gear

Ocean Optics put spectroscopy on a circuit board in 1992.

Descendants of that box are on Curiosity right now.

The expensive instrument rack doesn't die when someone builds a bigger lab. It dies when someone needs the measurement more than they need the building.

[Alex Towfigh @alextowfigh]

I spent 6 years building water sensors.

Then we analyzed 7.17M data points from a wastewater plant and found a pump starting every 25 seconds, its backup doing 2x the work, sensors disagreeing by 8x, and a critical dead tag.

The plant didn’t need more data. It needed something connecting the dots.

That’s when it clicked: the bigger opportunity in industrial AI isn’t another sensor or dashboard.

It’s a system that continuously detects anomalies, figures out why they’re happening, what they’ll cost, what to do next, and whether the fix worked.

That’s what we’re building with AquaMesh.

[Alex Towfigh @alextowfigh]

@Teo_Sinamin Makes sense. Since China dropped mandatory storage for grid hookups in 2025, long duration has to earn its keep, so the hybrid is a smart hedge.

[Alex Towfigh @alextowfigh]

That humanoid robot restocking shelves at night is a person in a VR headset.

The smart part: the store pays per shift, so the customer funds the training data.

Most "autonomous" robots launch this way. The question is when the headset comes off.

[CyberRobo @CyberRobooo]

The night shift nobody wants, now a training ground for humanoids

At a Føtex supermarket near Copenhagen, a Unitree G1 is learning that job. Danish start-up Hive Robots has been training them since June with Salling Group, Denmark's largest retailer.

In the footage, a person in a VR headset guides every move. Each can and bag of pasta teaches the AI something. The ambition is restocking on their own by early 2027.

Hive doesn't sell the robot. Stores book working hours: one robot, one teleoperator, 8 hours a night once the doors close, paid by the shift. Next step on their roadmap is three robots per store.

Føtex says staff spend 40 to 50% of their time refilling shelves. Hand that over and people get time back for customers. It's one store for now, but Salling runs 2,000+ stores across six European countries.

Still slower than people, still closely supervised, and Danish law would have to change before they work among shoppers.

The store is closed. Someone still has to refill thousands of shelves before morning. Let the humanoids take that shift.

[Alex Towfigh @alextowfigh]

@jdh For hardware founders it's earned in one moment: the bridge round nobody else will fund. That's an option, not a stream of value.